7 posts
Trusting AI output gives you tests that can't fail and an assistant that invents policy. Review every AI-written test by breaking the feature on purpose, keep agents out of the gate, and test AI features with eval sets run in CI.
Tests that run only when someone remembers put the gate in a person. Run them in four CI tiers that each block the next, keep the commit gate under ten minutes, give every run fresh data, and quarantine flaky tests the same day.
A green functional suite can't see a renamed field, a slow peak, an open security hole or a form a keyboard can't finish. Add a Pact contract, a k6 test against an SLO, a ZAP baseline scan, and axe plus a keyboard pass.
Flaky UI tests teach a team to ignore red. Fix the three usual causes: brittle locators, fixed sleeps and shared data. Use role and test-id locators, web-first assertions, API setup, Accessibility IDs on mobile, and same-day quarantine.
Testing only through the UI is slow and misses what anyone can do with curl. Write your first pytest tests against the API: the €50.00 boundary, the order chain from login to delete, and the two-user test for broken object level authorisation.
Testing with whatever values come to mind misses the bug at the edge and leaves a coverage claim nobody can check. Pick values with partitions, boundaries and decision tables, explore in timed sessions, and write bug reports a developer can reproduce.
Learning testing tools in a random order leaves you with a suite that runs but can't say what it covers. Learn in order instead: test design, one language and the API, the UI, the checks functional tests miss, CI, then AI, all on one checkout.
